Carbon-A Maps 566 Million Predicted Coding Loci Across 22,617 Taxa

Carbon-A launched Oct. 8 with a database of 566 million predicted protein-coding loci, giving researchers candidates across 48,167 eukaryotic genome assemblies and 22,617 taxa for genome annotation.

The database spans about 27 trillion base pairs. Each entry includes an assembly, genomic coordinates, a reconstructed coding sequence, a predicted protein sequence and a confidence score. The records are candidates for further research, not 566 million experimentally confirmed discoveries.

Carbon-A has 1.2 billion parameters and analyzes DNA in windows of 98,304 base pairs. It predicts coding sequences at single-base resolution on both strands and is designed for mammals and other vertebrates, invertebrates, plants, fungi and protists. The model is licensed under Apache 2.0.

Genome annotation identifies genes and coding regions in sequenced genomes. Carbon-A is built for that task, while Carbon-500M, Carbon-3B and Carbon-8B are generative DNA models that predict DNA sequence tokens. The distinction separates finding possible coding regions in existing genomes from generating DNA sequences.

Across 42 benchmark genomes, Carbon-A scored 0.944 on macro-averaged nucleotide F1. It also outperformed the benchmark's comparison systems at the nucleotide, exon and gene levels.

RNA data from PacBio Iso-Seq experiments involving cats, Syrian hamsters, chickens and Arabidopsis supported some Carbon-A predictions that were absent from RefSeq, a reference genome database. The evidence supports transcription and exon structure. It does not establish that the sequences are translated into proteins or reveal what those proteins do.

At launch, about half of the target GenBank genomes had been annotated. The database's predicted loci therefore offer researchers a broad set of candidates to investigate, while the RNA evidence provides support for some predictions without confirming protein production or function.